Emotions are fundamental for players to deeply engage with games. Players’ responses in a game are affected by their emotional states which if, in turn, could affect the way the game responds the player-game interaction could be augmented and enriched by magnitudes realizing affective loop-enabled games. Games may evolve and adapt to the player in many different ways and convey emotions through a variety of techniques and effects. In this section we will discuss emotion adaptation and emotion expression, placing it in the context of the affective loop discussed earlier. The adaptation module of the affective loop should be able to provide satisfactory answers to – at least some – of the following questions: which stimulus (or playful experience) should be presented next? When should it be presented? Which game elements should be adjusted and how?
Arguably, we can achieve meaningful adaptation in games because players are prepared for personalised experiences more than in any other form of human computer interaction. The players’ relationship to game adaptation is dependent on their playing style, experience, personality etc. and the form of adaptation (e.g. implicitly or explicitly) needs to comply with the player needs. So, when creating and designing emotional games, one needs to consider all the processes involved, starting with the game design process itself. Further, while emotion models can be used to inform game designers in a mixed-initiative design fashion (see (Smith, et al., 2011), (Liapis, et al., 2012) among others) we argue that a semi- of fully-automated approach to emotion-driven game design can ultimately lead to improved playing experience. But as the game design entails the definition of many aspects of a game, when referring to emotional game adaptation one fundamental question to ask is what game elements one can adjust? In other words: what does emotional adaptation entail? A high-level observation of available game elements derives two key classes of adaptable game features: game agents (and NPCs) and game content (see Figure 1). Both of these can be manipulated to convey emotional responses and adaptation, in a manner that leads the player to become more emotionally involved with the game.
Adapting and Expressing Emotion through Agents and NPCs
One of the two main ways by which emotions can be manifested in games is through their game characters (see Figure 1). Characters in a game need to act, and their actions should be determined by emotional reactions to events occurring in the game. This can be achieved in a completely scripted manner, or through an automatic, autonomous approach, by using emotion agent architectures (Gratch & Marsella, 2004) underlying cognitive models to generate behaviour of the characters. Such architectures are usually model-based as they seek inspiration in psychological or physiological models of humans, and other species, and embed features that allow them to go beyond the pure “rational” behaviour. Emotional agent architectures naturally include a way to capture emotions or other affective states, such as moods or even personality (Doce, Dias, Prada, & Paiva, 2010). These affective states often have symbolic representations, or can be the resulting pattern of behaviour arising from a variety of different processes embedded in the agent. Examples of these architectures are EMA (Gratch & Marsella, 2004) and FAtiMA, used for research on serious games in the areas of social and emotional training (Paiva, et al., 2004), (Aylett, et al., 2009), (Lim, Dias, Aylett, & Paiva, 2012), ALMA (Gebhard, 2005), or the MindModule (Eladhari & Mateas, 2008) for player characters. Further, these characters may portray social roles and have different personalities leading the users to raise expectations concerning the characters actions, and as such triggering emotional reactions by the players when those expectations are not met. A game character that plays an ally or a mentor (see (Isbister K. , 2006)) will lead to certain emotional reactions when for example the character deceives the player. The personality of a game character can be established by the nature and strength of the emotions that the character portrays in different situations, and its tendency to act in a certain manner. For example, an extrovert character will use more speech acts and more expressive actions than an introvert character. These features of personality may be achieved by the appropriate parameterization of the agents (see (Doce, Dias, Prada, & Paiva, 2010)). Characters will not only trigger emotional states as a response to a given situation, but they also need to express emotions in a way that conveys their “internal” emotional state. Thus, emotions not only guide the decision making of the characters, but also the expressions they will portray, which again can be generated in an automatic manner. Expressions of different emotional states, such as for example fear, surprise, sadness or happiness may blend handcrafted animations to express both strong and subtle emotions with procedural animation techniques to achieve real-time behaviour animated characters (Perlin & Goldberg, 1996). Characters provide a rich medium to express emotions, trigger emotions and adapt to the emotions of players. Further, these emotional manifestations can be augmented via adaptive narrative and camera profiles (Picardi, Burelli, & Yannakakis, 2011) allowing for the emphasis on particular emotional states or features, and combining it with game content adaptation (see Figure 1). We should, however, stress the research oriented nature of these early systems acknowledging that autonomous emotional NPCs are still in the realm of a few exploratory research projects. However, we believe that by addressing this challenge, this area will become one of the major pillars of AI in games.
Adapting and Expressing Emotion through Game Content
Yet, games may or may not include agents. Games, however, definitely include a form of virtual environment where agents “live” (if existent) and the interaction is taking place. There are a number of elements (i.e. game content) from the game world that an adaptive process can alter in order to drive the player to particular affective patterns. As mentioned already, game content may include every aspect of the game design such as game rules (Togelius & Schmidhuber, 2008), reward systems, lighting (de Melo & Paiva, 2007), camera profiles (Yannakakis, Martinez, & Jhala, 2010), maps (Togelius, et al., 2010), levels, tracks (Togelius, Yannakakis, Stanley, & Browne, 2011), story plot points (Riedl, 2012), and music (Eladhari, Nieuwdorp, & Fridenfalk, 2006). Even behavioural patterns of NPCs such as their navigation meshes, their parameterised action space and their animations can be viewed as content.
The adaptive process in this case is referred to as procedural content generation (PCG) which is the generation of game content via the use of algorithmic means. According to the taxonomy presented in (Togelius, Yannakakis, Stanley, & Browne, 2011) game content can be necessary (e.g. game rules) or optional (e.g. trees in a level or flying birds on the background). Further, PCG can be either offline or online, random or based on a parameterised space, stochastic or deterministic and finally it can be either constructive (i.e. content is generated once) or generate-and-test (i.e. content is generated and tested). The Experience-driven PCG framework (Yannakakis & Togelius, 2011) views game content as an indirect building block of player affect and proposes adaptive mechanisms for synthesizing personalised game experiences.
Integration in the Affective Loop: When and How to Adapt
Once sufficient amounts of appropriate game stimuli (which include the actions of the game characters and in the environment) have been presented to the player, aspects of the playing experience can be detected and modelled. For the affective loop to close effectively the game logic needs to adapt to the current state of the game-player interaction. Whether agent behaviour or parameterised game content, a mapping is required linking a user’s affective state to the game context. That mapping is available as it is essentially the outcome of the emotion modelling phase. Any search algorithm (varying from local and global search to exhaustive search) is applicable for searching in the parameterised search space and finding particular game states (context) that are appropriate for a particular affective state of a specific player. For example, one can envisage the optimization of agent behaviour attributes for maximizing engagement, frustration or empathy towards a player (Leite, et al., 2010). As another example, the study of Shaker et al. (2010) presents the application of exhaustive search for generating Super Mario Bros (Nintendo, 1985) levels that are maximally frustrating, engaging or challenging for any player. In that study parameterised game levels are linked to in-game player behaviour attributes and a set of affective states inferred from crowdsourced player reports. The model-free affective model is built via evolving neural networks that learn the crowdsourced pairwise preferences (i.e. neuro-evolutionary preference learning) .
A critical question once an adaptation mechanism is designed is how often particular attributes should be adjusted. The frequency can vary from simple pre-determined or dynamic time windows (Yannakakis & Hallam, 2009) but adaptation can also be activated every time a new level (Shaker, Yannakakis, & Togelius, 2010) or a new game (Yannakakis & and Hallam, 2007) starts, or even after a set of critical player actions – such as in Façade (Mateas & Stern, 2003). The time window of adaptation is heavily dependent on the game under examination and the desires of the game designer. Regardless of the time window adopted, adaptation needs to be interwoven well with design if is to be successful.
One approach for assessing the appropriate time window for game adaptation is to test the validity of the emotion models in different time windows and then make a compromise between adaptation frequency and model performance (Yannakakis & Hallam, 2009)). As models are expected to yield lower accuracies the more deviant they are from the interaction time window they were built on, one needs to evaluate their accuracy with respect to different time windows. A good compromise between accuracy and performance would yield sensible decisions about the length of the adaptation time windows. In general, those can be either static across all gameplay or dynamic (dependent on e.g. different levels)
